Research On Key Techniques For Human Modeling Based Markerless Motion Capture

For several decades, Markerless Human Motion Capture is a hot issue in computer vision area, for its extensive application value and academic value. In recent years, with the rapid development of computer software and hardware, research in Human Motion Capture is no longer confined to the field of human pose estimation, which extended to capture human shape model from a sequence of images or video, also called surface capture. Human motion capture datas have been used in the films and computer games widely. At the same time, the3D human model corresponding to each image frame is widespread in films, cartoons, computer games and industrial design also.In this thesis, we around the topic of human modeling based markerless motion capture to work, and research its key techniques. We focus on three important issues of this topic that how to capture3D human shape model from a multi-view image sequence, how to build parameterized human model and use it to generate human model used in human motion capture initialization step, how to do human motion capture using local optimization and global optimization algorithm based a detail3D human model. The main contributions of this thesis are summarized as follows:1. We propose a new frame work of human template model based human Pose and Shape Model Adaptation step by step, called PSMA. In this frame, the first step is to make the template model pose to fit body pose in images. Then, do the human shape fitting step. Due to human shape change will cause human pose change a little. In order to get the correct result, we need to optimize human pose again. The template human model is obtained by the first frame image, which shows clear body parts. The pose deformation can be controlled using the embed skeleton. For the body shape deformation, we use the Laplacian deformation technology to realize shape deformation and match with the body silhouettes in images. This method can be used in the capture scene that human with common clothes or loose clothes. It can resolve the problem of the human model topology error which is builded using3D stereo vison reconstruction algorithm in dynamical human modeling area, and through experiments confirmed its feasibility.2. We propose a new method of estimate the human pose and shape parameters using the human parameterized model SCAPE. We build the popular SCAPE model using the human scan data base. Using the SCAPE model in the PSMA frame, and generate a3D human model for motion capture initialization step. The pose deformation can be controlled using the embed skeleton also, using the fit errors between human model projection silhouettes and body silhouettes in images. According to optimize a function of pixel distance to estimation human shape parameters and pose parameters. And we can get the3D human model is consistent with the target object. We test our method using input images, using the generated3D human model compare with the3D scan model and3D reconstruction human model, which can prove our method is valid.3. We propose a new energy function according to human pose prior knowledge constraints. Based on this function, we combine the local and global optimization algorithms to estimate human pose from multi-view images. It not only use human body silhouettes fitting, but also fitting the texture features of neighbour frames to improve the accuracy of local optimization algorithm. We use the energy function to constraint the particles, and reduce the number of particles and iterations, improve the efficiency of the whole estimation process. And using the3D detailed human model, compared with the stick human model, it can improve the pose estimation result also.